clustershell
clustershell copied to clipboard
cea-hpc
Performance on large many-dimensional nodesets
HPE Cray EX supercomputers use hardware locations (called xnames) that encode up to 5 dimensions. An example compute node might be x1000c2s3b0n1. We don't use xnames for our compute nodes, but our switches, bmcs, chassis controllers, etc do use them. Some of our local tooling (and clush/cluset) struggle with long lists of xnames, particularly when folding.
[[email protected] ~]# time (cluset -e x[2000-2073]c[0-7]s[0-7]b[0-1] | cluset -f)
x[2000-2073]c[0-7]s[0-7]b[0-1]
real 6m38.483s
user 6m38.228s
sys 0m0.088s
Any thoughts on how we could improve this? Thanks
Thanks for the report @mattaezell.
A quick look (on the master branch) shows that most of the time is spent in RangeSetND._fold_multivariate_merge() which does the nD folding:
$ cluset -e x[2000-2073]c[0-7]s[0-7]b[0-1] | python3 -m cProfile -s cumulative lib/ClusterShell/CLI/Nodeset.py -f
x[2000-2073]c[0-7]s[0-7]b[0-1]
1225113449 function calls (1073087573 primitive calls) in 848.288 seconds
Ordered by: cumulative time
ncalls tottime percall cumtime percall filename:lineno(function)
107/1 0.009 0.000 848.388 848.388 {built-in method exec}
1 0.000 0.000 848.388 848.388 Nodeset.py:27(<module>)
1 0.001 0.001 848.258 848.258 Nodeset.py:334(main)
1 0.000 0.000 848.257 848.257 Nodeset.py:155(nodeset)
1 0.251 0.251 848.255 848.255 Nodeset.py:44(process_stdin)
18950 0.028 0.000 846.483 0.045 RangeSet.py:884(inner)
152119181/94356 14.296 0.000 845.921 0.009 {built-in method len}
9473 0.008 0.000 845.920 0.089 RangeSet.py:1147(_fold)
3 0.000 0.000 845.911 281.970 NodeSet.py:238(__len__)
3 0.000 0.000 845.911 281.970 RangeSet.py:926(__len__)
1 0.000 0.000 845.856 845.856 RangeSet.py:1180(_fold_multivariate)
1 186.997 186.997 844.685 844.685 RangeSet.py:1245(_fold_multivariate_merge) <<<
71670778 86.067 0.000 333.553 0.000 RangeSet.py:533(copy)
47701583 31.013 0.000 331.814 0.000 RangeSet.py:582(__and__)
47701583 33.374 0.000 287.534 0.000 RangeSet.py:591(intersection)
71746554 115.193 0.000 176.797 0.000 RangeSet.py:106(__init__)
23855527 18.540 0.000 173.677 0.000 RangeSet.py:564(__or__)
23855527 16.714 0.000 147.369 0.000 RangeSet.py:573(union)
51414608 57.401 0.000 144.083 0.000 RangeSet.py:542(__eq__)
414319969 110.208 0.000 110.208 0.000 {built-in method isinstance}
95526306 67.220 0.000 91.474 0.000 RangeSet.py:736(update)
50787459 38.299 0.000 72.265 0.000 RangeSet.py:652(issubset)
52033619 17.782 0.000 34.702 0.000 RangeSet.py:676(_binary_sanity_check)
47701583 31.294 0.000 31.294 0.000 RangeSet.py:703(intersection_update)
71784438 19.991 0.000 19.991 0.000 RangeSet.py:244(set_autostep)
623080 0.775 0.000 2.092 0.000 RangeSet.py:670(__gt__)
18947 0.021 0.000 1.969 0.000 NodeSet.py:1508(update)
9474 0.041 0.000 1.760 0.000 NodeSet.py:1202(__init__)
18947 0.031 0.000 1.339 0.000 NodeSet.py:789(parse)
9472 0.061 0.000 1.296 0.000 NodeSet.py:810(parse_string)
28419 0.043 0.000 0.982 0.000 NodeSet.py:539(update)
37889 0.074 0.000 0.959 0.000 NodeSet.py:490(_add)
623080 0.481 0.000 0.809 0.000 RangeSet.py:657(issuperset)
18944 0.042 0.000 0.735 0.000 NodeSet.py:996(_scan_string)
2 0.000 0.000 0.640 0.320 RangeSet.py:1126(_sort)
2 0.061 0.031 0.640 0.320 {method 'sort' of 'list' objects}
9472 0.190 0.000 0.584 0.000 NodeSet.py:962(_scan_string_single)
9473 0.033 0.000 0.579 0.000 RangeSet.py:1128(rgveckeyfunc)
18946 0.039 0.000 0.534 0.000 RangeSet.py:898(copy)
1 0.011 0.011 0.531 0.531 RangeSet.py:1190(_fold_multivariate_expand)
75776 0.154 0.000 0.526 0.000 RangeSet.py:194(fromone)
We'll investigate.
Thanks for the super-quick patch for this. Testing seems to work MUCH better.
[[email protected] ~]# time (cluset -e x[2000-2073]c[0-7]s[0-7]b[0-1] | cluset -f)
x[2000-2073]c[0-7]s[0-7]b[0-1]
real 0m2.849s
user 0m2.956s
sys 0m0.084s